Battery Pack Generator · model check

Model vs Pack

The generator's thermal and electrical model, run over 30 real rides of an electric scooter and checked against what the pack's own sensors recorded in the same seconds. The model is given time, current and the pack's starting charge. Voltage and temperature are its predictions.

30 rides
the data
123.7 km, 13.9 h recorded, odometer 105 → 223.4 km, 21 July — 19 August 2026; runs under 1 km excluded
0.56 V
voltage
median RMSE against the pack's own voltmeter; 0.37 V of it is shape, the rest a constant offset
0.87 °C
pack temperature
median RMSE; on rides under 20 minutes, 0.45 °C
6.8×
what the new input bought
on the 20 rides that did not start full, median voltage RMSE fell 3.53 → 0.52 V once v1.36.0 could be told the starting charge

how it was run

Run conditions

Two numbers come from independent measurements on this scooter; everything else is an application default. No parameter was tuned against the curves being compared.

pack
14S7P, 98 × 18650 2500 mAh, the app's own preset
cell resistance
64 mΩ → packR 130 mΩ, to match the 128 mΩ measured on the pack
cooling
1.5 W/K — the default; three cooldowns measured 1.46
start charge
per ride, what the BMS coulomb counter read when it began — 30 % to 100 %, median 86 %. The field arrived in v1.36.0; this report is the first run to use it
ambient
per ride, the pack's own reading at its start: the model has no separate field for a starting temperature
current in
current_in from the controller at 4 Hz — the model's integration step is the same
checked against
v_in and the mean of the two sensors sitting on cells (temp1, temp3)
sources
7 dumps from dumps/ and 3 rebuilt from logs/ out of their LOG BLK lines
deduplication
by overlapping records: the VESC odometer rolls back after a power cycle and cannot be a key
runs per ride
two — through importPoints() as a user gets it, and with the full trace
selection
runs shorter than 1 km excluded: the pack does not warm up in them

Where the two resistances come from. The 128 mΩ is measured, not chosen: the scooter repo confirms it three independent ways — v_in against current_in on the controller side gives 128 mΩ, pack_v against pack_i on the BMS side 129, and the fourteen groups summed separately 128.7.

The 64 mΩ is then back-solved rather than measured. In a 14S7P pack the cells contribute s/p = 14/7 = twice one cell's resistance, so 128/2 = 64 is what the field needs to hold for the model's total to land on the measured one. It happens to agree with the repo's own per-cell estimate of about 62 mΩ — but that estimate carries a warning worth repeating: it includes the interconnects, the connectors and the BMS FETs, and calling it cell resistance would be wrong. A healthy 18650 of this type sits nearer 30–40 mΩ, so a third to a half of the 64 is something other than the cell itself. This log cannot separate the two.

For what is compared here, it does not have to. Voltage is predicted at the controller, where the sag is the current times the whole series path, so one lumped resistance is the right construction rather than an approximation. And simulate() computes heat as I²·packR against a heat capacity of cells only, so it puts all of the heat into the cells however the resistance is attributed — splitting the same 128 mΩ between cells and joint resistance would not move the temperature curve at all. The split would matter to the per-element heat map, which nothing here tests.

one ride, whole

15.71 km, 115 minutes

The longest record in the archive. On top, the current the model was given; below it, what the model made of that, over what the pack measured. The stairs on the measured temperature are the sensor's 1 °C step, not noise.

where the model drifts

Error against duration

One point per ride. Temperature carries the only systematic dependency in the set: the longer the record, the worse it gets. Voltage shows no such thing.

temperature RMSE, °C
voltage RMSE without offset, V

all 30

Temperature: predicted and measured

In odometer order. Each card carries its own vertical scale — what is being compared is the shape of the curves, not absolute levels between rides.

measured model

all 30

Voltage: predicted and measured

The same rides in the same order. Start charge given is the model as it stands: told what the pack actually held, it lands on the measured curve closely enough that the blue line disappears under the orange one on most cards. Ignored is the same run with the field back at its 100 % default — every prediction before v1.36.0 looked like that, and on a ride begun half empty the whole curve rides volts high.

measured model

per ride

Full table

“was” is the voltage RMSE the same ride produced before the starting-charge input existed; “no offs.” is today's RMSE after its remaining constant offset is subtracted; “rise” is the temperature climb over the ride, model against measurement.

odometerkmminstartrms current V RMSEwasV offs.V no offs. T RMSErise mod.rise meas. pointsverdictsource

what follows from it

Findings

Voltage: 0.56 V, and most of what is left is the OCV curve

Median RMSE against the pack's own voltmeter is 0.56 V on a signal that sags 6–7 V under load. Of that, 0.37 V is shape — the part current-times-resistance has to get right — and the remainder is a constant offset that now spans −1.27 to +0.62 V instead of the −1.27 to +8.09 V it spanned before.

The rides that still carry the most error are, inverted from before, the ones that started full: their median RMSE is 1.04 V and did not move, because there was nothing about their starting charge to correct. What they are showing is the OCV curve itself — a generic NMC table against this particular pack, running about a volt low at the top. That is now the largest single term, and the next one worth attacking.

What the starting-charge input bought

Before v1.36.0 the model began every run at 100 % charge, because there was nowhere to say otherwise. Across these 30 rides that cost:

Median voltage RMSE 1.46 → 0.56 V. p90 4.56 → 1.07 V. Worst ride 8.10 → 1.33 V. On the 20 rides that did not begin on a full pack, median 3.53 → 0.52 V — a factor of 6.8. On the 10 that did, 1.04 → 1.04 V, unchanged, which is the control: a correction that also “improved” those would have been a fit rather than a fix.

Temperature did not move at all, and could not have: heat is I²·packR, with no term containing charge anywhere in it.

Heating is predicted more precisely than it is measured

The median temperature error is 0.87 °C, and 0.45 °C on rides under twenty minutes. The sensors step in whole degrees, which is σ ≈ 0.29 °C on its own, so on short rides the model is up against the resolution of the measurement rather than its own limits. The median error in temperature rise over a ride is +0.34 °C — and neither cell mass nor heat capacity was tuned to this data.

The error accumulates from riding, not from standing

Under 20 minutes the median is 0.45 °C; longer, 1.62 °C, reaching 4.2 °C. Movement is what builds it: the error correlates with time under current at 0.56, with total duration at 0.48, and with the share spent parked at −0.11 — standing still, if anything, dilutes it. The stops are not incidental: the median ride is stationary for 42 % of its record.

What grows is the scatter, not a drift. Time under current barely relates to the error in temperature rise (−0.03), and on the longest rides that error goes both ways, from −2.6 to +4.3 °C. That points at ambient: one field serves as both the starting temperature and the temperature of the surroundings, they cannot be set apart, so a run begins at the pack's own reading and cools back to it. How wrong that is, and in which direction, is settled by the particular day. The numbers agree: the temperature offset correlates with the product of the ambient error and the duration at 0.54 — higher than with either factor alone (0.47 and 0.16).

A zero while parked is a reading, not a gap

While the scooter stands, both current sensors write exactly 0.000, and write it as a value: there are no missing samples in those seconds, and voltage keeps reporting live in the same rows. So nothing was filled in with zeros — not in the log, and not in the data preparation, where a missing sample is dropped rather than turned into one. The BMS shunt steps in 0.01 A, and the standing draw does not reach even that.

None of the thermal conclusions turn on it: a hundred milliamps through 130 mΩ is milliwatts. What it does mean is that self-discharge and parasitic draw while parked are invisible in the amp-hour accounting, and this log cannot recover them.

Not everything is checked. All of the above concerns the thermal and electrical model on the scale of minutes and hours. The steady-state temperature the app's verdict rests on is not tested by this data at all: about 6 % of the heat generated over a ride leaves the pack, and the cooling coefficient barely participates in a balance like that. The geometry — cell packing, the waist of the interconnects, busbar section, current sharing between cells — is tested by nothing here: the onboard log knows nothing about it.

The import thinning costs less than it looked

The median ride is 2178 points, which importPoints() compresses to 70. At the median that costs 3.1 % of the rms current, and 9.3 % at worst. Short active stretches lose more than long ones: their peaks have less to average against.

The two-hour ceiling cuts part of the archive off

The profile duration field stops at 7200 s, and beyond it the app enumerates no configurations at all. Four rides had to be cut, and the window has to be anchored on the first movement: in sessions that open with hours of standing still, the first two hours of record contain no riding.